Discussion Intel current and future Lakes & Rapids thread

[cbn]

No what I mean is you need Optane to be paired with at least equal number of DIMMs. So you can have 6 DIMMs and 3 DC PMMs but not 3 DIMMs and 6 DC PMMs.

Yes, I did know that was what you were referring to.

Sorry for the confusion.

I have should have wrote despite Xeon-AP being not being able to pair DRAM DIMMs 1:1 with Optane DC PMM it is still possible to achieve the Intel recommended ratio of 1GB DRAM to 8GB Optane if enough HBM2 is present.

 

[repoman27]

Intel is in the process of moving almost all of the capacity for 10 nm back to 14, if they haven't already completed that.

What capacity for 10 nm? D1D and Qiryat Gat appear to be the only fabs with P1274 tooling and we know that Fab 28 has also been producing 14 nm for probably at least a year now. Allegedly Israel is still making Ice Lake though.

Apple is typically late with product introductions. If they wanted a January introduction, then Icelake would make sense.

Also, despite Ivy Bridge U chips being introduced June at Computex, I had to wait until November to get my XPS 12. After waiting a month for shipping, I had it on my hands in January.

Both Intel and the OEMs would ideally like to achieve as near as possible a 12 month release cadence to maximize upgrade cycles and drive revenue. In principle, Apple wants to release Mac hardware refreshes within a few weeks of Intel being able to deliver them sufficient quantities of chips, but sometimes delays happen for reasons. They never arbitrarily wait to ship, and enjoy leading whenever they can (like with Xeon W for the iMac Pro and new Mac Pro). In the case of the MacBook Air and 13-inch MacBook Pro, Apple isn't waiting for an alignment of the stars (CPU, PCH, and Thunderbolt controllers from Intel, GPU from AMD, SSD controller from Samsung, T-series chip to come back from TSMC) or introducing a new chassis design. Furthermore, the new MacBook Air chassis was essentially custom designed for Ice Lake, and now Apple's options are either to be late to the party with Ice Lake or release another update less than 6 months after the current one. The MacBook Air is ostensibly Apple's top selling Mac model, and right now it only has a single Amber Lake-Y CPU option—that's embarrassing, and Apple has to be furious about the situation.

I think jpiniero got it right: Intel's 10 nm still isn't even yielding enough to produce Y and U series chips in volume.

 

[jpiniero]

What capacity for 10 nm? D1D and Qiryat Gat appear to be the only fabs with P1274 tooling and we know that Fab 28 has also been producing 14 nm for probably at least a year now. Allegedly Israel is still making Ice Lake though.

Intel had several 10 nm production lines ready to go at one point. What you mentioned might be what's left.

 

[Dayman1225]

Looks like Tremont may introduce L3 cache for Atom

https://twitter.com/x/status/1150444073821413376

 

[IntelUser2000]

Both Intel and the OEMs would ideally like to achieve as near as possible a 12 month release cadence to maximize upgrade cycles and drive revenue. In principle, Apple wants to release Mac hardware refreshes within a few weeks of Intel being able to deliver them sufficient quantities of chips, but sometimes delays happen for reasons.

Actually I saw it few times in the past where Intel wanted it every 12 months but had to delay it slightly because manufacturers wanted it slightly longer than 12 months. Apparently, a strict 12 month cycle can be suboptimal for sales.

And what I mean by Apple is that Icelake is likely really a holiday product and that's several months off the cycles they want to introduce. You might see a model or two in August, but most should be November/December.

 

[Dayman1225]

I wonder what process Intels discrete GPU next year is based on.

Going back to this, this slide implies 10nm for 2020 GP-GPU

 

[Ajay]

Uh, I thought GP-GPU (Xe) was going to be on 7nm!??

 

[Dayman1225]

Uh, I thought GP-GPU (Xe) was going to be on 7nm!??

That's the 2021 Iteration.

 

[IntelUser2000]

Uh, I thought GP-GPU (Xe) was going to be on 7nm!??

Yea, Tigerlake uses the new Xe architecture, and should also be in dGPUs. Then the next iteration of Xe becomes the lead product for 7nm on datacenter.

So about a year from now we might see Intel dGPUs. Should be very exciting in the GPU market with 3 players all launching next generation products!

 

[ondma]

Yea, Tigerlake uses the new Xe architecture, and should also be in dGPUs. Then the next iteration of Xe becomes the lead product for 7nm on datacenter.

So about a year from now we might see Intel dGPUs. Should be very exciting in the GPU market with 3 players all launching next generation products!

Lets hope so, but with all the problems they are having with their cpus, I am not expecting much from the dgpus. In fact, I am afraid they are going to flop in that area just like they did in mobile. Maybe bring out a few non-competitive products and then drop it. If they still had a process lead, I would be more optimistic, but they dont, and both AMD and nVidia have a lot more experience in dgpus.

 

[Ajay]

What fabs are ramping up 7nm, other than fab 42 @ Chandler, AZ?

 

[Dayman1225]

What fabs are ramping up 7nm, other than fab 42 @ Chandler, AZ?

Probably Ireland fab. No official comment from Intel

 

[liahos1]

Question for someone just getting back into the flow. Is sapphire rapids based on 7nm xeon?

 

[cbn]

If they still had a process lead, I would be more optimistic, but they dont, and both AMD and nVidia have a lot more experience in dgpus.

Don't forget about ODI.

This should allow a GPU built with many small or medium size dies to operate essentially as a "larger than reticle" monolithic die.

 

[Dayman1225]

Question for someone just getting back into the flow. Is sapphire rapids based on 7nm xeon?

Nope, 7nm Xeon is Granite Rapids.

 

[Dayman1225]

Don't forget about ODI.

This should allow a GPU built with many small or medium size dies to operate essentially as a "larger than reticle" monolithic die.

That appears to be the plan with their 2021 GP-GPU that will be powering Aurora21 Super Computer

 

[Ajay]

Don't forget about ODI.

This should allow a GPU built with many small or medium size dies to operate essentially as a "larger than reticle" monolithic die.

That'll work great for compute tasks. Big question is, how to do it for GFX without killing performance.

 

[Dayman1225]

That'll work great for compute tasks. Big question is, how to do it for GFX without killing performance.

Probably sucks. Client dGPU will likely continue to be monolithic for now

 

[IntelUser2000]

That'll work great for compute tasks. Big question is, how to do it for GFX without killing performance.

The issues with interconnects are problems in compute too. It's a huge thing in the HPC world. That's why they focus so much on it. Especially if you want thousands of them, even 1% loss in scalability will reduce performance of the end system by a significant amount.

With client systems games are moving away from Crossfire and SLI setups. Game developers can make it work, but its their choice to do so and supporting it adds to development time. It's the latency and reduction in bandwidth introduced by off-die solution that causes all the nasty micro-stutter and compatibility issues both games and devs don't want to deal with.

Nvidia with CUDA is the defacto standard in GPU compute world. Software is the ultimate differentiator. Once the momentum is there, it doesn't matter if your hardware is better. The software development work already went through all the little details(like dealing with scalability issues on a particular GPU architecture), and they don't want to do it again with another vendor.

There's simply no guarantee Intel will even reach AMD levels of volume share in GPU compute. Even among x86 CPUs, the software still need optimization to maximize performance out of them. Nevermind in GPUs where the architecture is completely different.

I think Intel has a better chance of succeeding in client dGPUs if they have a good product. At least it has a big user base already. If it scales well, and they put in a lot of effort into software development, and has great perf/watt/$, sure people will go for it. It won't be an overnight success. Things never are. But it can be enough to build the basics to make it one.

This thread isn't about GPUs though. We can move the discussion there.

 

[birdie]

 

[jpiniero]

View attachment 8501

That's Comet Lake. Intel's marketing for Icelake vs Comet Lake is going to be something.

 

[DrMrLordX]

Question for someone just getting back into the flow. Is sapphire rapids based on 7nm xeon?

Don't expect 7nm anything from Intel until 2022. Except for datacenter Xe.

 

[IntelUser2000]

That's Comet Lake. Intel's marketing for Icelake vs Comet Lake is going to be something.

Comet Lake: Faster CPU, aimed at dGPU devices, up to 6 cores
Ice Lake: Faster GPU, better battery life, more compact, for iGPU systems

Comet Lake will likely end up in cheaper and larger notebooks as well.

 

[Ajay]

Don't expect 7nm anything from Intel until 2022. Except for datacenter Xe.

Source? If true, this would totally suck for Intel.

 

[Dayman1225]

Source? If true, this would totally suck for Intel.

Xe in 2021 is their lead 7nm product and Navin Shenoy, head of DCG, said that 7nm DC CPU would come after the GP-GPU, so in 2022 - Granite Rapids - as for Desktop? Not sure, could see Meteor Lake in 2022 on 7nm.